Question

How is end of file detected for a file in nonblocking mode?

Answer 1

At least on POSIX (including Linux), the obvious answer is that nonblocking regular files don’t exist. Regular files ALWAYS block, and O_NONBLOCK is silently ignored.

Similarly, poll()/select() et al. will always tell you that a fd pointing to a regular file is ready for I/O, regardless of whether the data is ready in the page cache or still on disk (mostly relevant for reading).

EDIT And, since O_NONBLOCK is a no-op for regular files, a read() on a regular file will never set errno to EAGAIN, contrary to what another answer to this question claims.

EDIT2 References:

From the POSIX (p)select() specification: “File descriptors associated with regular files shall always select true for ready to read, ready to write, and error conditions.”

From the POSIX poll() specification: “Regular files shall always poll TRUE for reading and writing.”

The above suffices to imply that while perhaps not strictly prohibited, non-blocking regular files doesn’t make sense as there would be no way to poll them except busy-waiting.

Beyond the above, there is at least some circumstantial evidence

From the POSIX open() specification: The behavior for file descriptors referring to pipes, block special files, and character special files is defined. “Otherwise, the behavior of O_NONBLOCK is unspecified.”

Some related links:

http://tinyclouds.org/iocp-links.html

http://www.remlab.net/op/nonblock.shtml

http://davmac.org/davpage/linux/async-io.html

And, even here on stackoverflow:

Can regular file reading benefited from nonblocking-IO?

As the answer by R. points out, due to how page caching works, non-blocking for regular files is not very easily defined. E.g. what if by some mechanism you find out that data is ready for reading in the page cache, and then before you read it the kernel decides to kick that page out of cache due to memory pressure? It’s different for things like sockets and pipes, because correctness requires that data is not discarded just like that.

Also, how would you select/poll for a seekable file descriptor? You’d need some new API that supported specifying which byte range in the file you’re interested in. And the kernel implementation of that API would tie in to the VM system, as it would need to prevent the pages you’re interested in from being kicked out. Which would imply that those pages would count against the process locked pages limit (see ulimit -l) in order to prevent a DOS. And, when would those pages be unlocked? And so on.